Detecting a problem with the nation’s power grid can be like looking for a needle in a massive haystack. Hundreds of thousands of interconnected sensors spanning across the United States capture real-time data on electric current, voltage, and other critical information, frequently taking multiple recordings per second.
Researchers at the MIT-IBM Watson AI Lab have developed a computationally effective method for automatically detecting anomalies in these data streams in real-time. They exhibited that their artificial intelligence technique, which learns to model the interrelation of the power grid, is far more effective than other popular techniques in identifying these defects.
Since the machine-learning model created by them does not need annotated data on power grid abnormalities for training, it would be easier for applying in real-world scenarios where high-quality, labeled datasets are frequently difficult to come by.
The model is also adaptable, and it can be used in other situations where a large number of interdependent sensors gather and report data, such as traffic monitoring systems. For instance, it could reveal traffic bottlenecks or how traffic jams spread.
In the case of a power grid, people have attempted in capturing the data by utilizing statistics and then defining identification rules with domain knowledge to say, for instance, when the voltage surges by a particular percentage, the grid operator should be warned. Even when aided by statistical data analysis, such rule-based systems need a significant amount of labor and expertise.
Senior author Jie Chen, research personnel and manager of the Watson AI Lab – MIT-IBM states that this process can be automated while also learning patterns from the data by utilizing advanced machine-learning methods, we can automate this process while also learning patterns from the data.
Enyan Dai – MIT-IBM Watson AI Lab trainee and graduate disciple at Pennsylvania State University, is one of the co-authors.
Exploring the probabilities
The researchers started by defining an abnormality as a low-probability event, such as a sudden spike in voltage. The power grid data was treated as a probability distribution, so if they were able to evaluate the probability densities, they will be able to identify the dataset’s low-density values. Anomalies correspond to data points that are least likely to occur.
Estimating those probabilities is difficult, especially because every sample records multiple time series, and every time series is a collection of multifaceted data points recorded over time. Furthermore, the sensors capturing all of that data are relying on one another, which means they are linked in a specific configuration and one sensor can sometimes have an impact on others.
The researchers utilized a particular type of deep-learning model known as a normalizing flow for learning the complicated conditional probability distribution of the data, which is especially productive at evaluating the probability density of a sample.
The normalizing flow model was supplemented with a Bayesian network, a kind of graph that can master the intricate, causal relationship structure among different sensors. According to Chen, this graph structure allows researchers for viewing patterns in the data and estimating abnormalities more accurately.
He states that the sensors interact with one another, and they have causal relationships and rely on one another. As a result, we must be able to introduce this dependency information into the way probabilities are computed.
This Bayesian network factorizes, or deconstructs, the joint likelihood of the multiple time series data into less complicated, conditional probabilities that are simple for parameterizing, learning, and evaluating. This permits the researchers for estimating the probability of monitoring particular sensor readings, and for identifying those readings that contain a low probability of materializing, meaning they are abnormalities.
Their method is particularly effective since this complicated graph structure need not be predefined — the model can master the graph on its own, unsupervised.
A significant technique
They put this framework to the test by seeing how well it could detect abnormalities in a power grid, traffic, and water system data. The datasets they utilized for testing involved human-identified abnormalities, allowing the researchers for comparing the abnormalities their model detected with real defects in each system.
By identifying a higher percentage of actual abnormalities in each dataset, their model surpassed all baselines.
Chen states that a lot of baselines don’t incorporate graph structure which completely authenticates our hypothesis. It’s assisting us to figure out the dependency relationships among the different nodes in the graph.
Their approach is also adaptable. Equipped with a large, unlabeled dataset, the model can be calibrated for making effective abnormality forecasts in other scenarios, like traffic patterns.
According to Chen once deployed, the model will continue learning from a stable stream of new sensor data, transforming to a possible shift in the data distribution and preserving precision over time.
Though this project is nearing completion, he is excited to apply the lessons learned to more areas of deep-learning research, especially graphs.
This approach could be utilized by Chen and his colleagues for establishing models that map another complex, conditional relationships. They also want to investigate how they can learn these models efficiently when the graphs become massive, possibly with millions or billions of interrelated nodes.
Instead of detecting abnormalities, this approach can be utilized for enhancing the precision of dataset-based forecasts or streamlining other categorization techniques.
The MIT-IBM Watson AI Lab and the United States Energy Department financed this research.
